Planning to Achieve Goals References and Utility

Goals, as typically conceived in AI planning, provide an insufficient basis for choice of action, and hence are deficient as the sole expression of an agent’s objectives. Decision-theoretic utilities offer a more adequate basis, yet lack many of the computational advantages of goals. We provide a preferential semantics for goals that grounds them in decision theory and preserves the validity of some, but not all, common goal operations performed in planning. This semantic account provides a criterion for verifying the design of goal-based planning strategies, thus providing a new framework for knowledge-level analysis of planning systems. Planning to achieve goals In the predominant AI planning paradigm, planners construct plans designed to produce states satisfying particular conditions called goals. Each goal represents a partition of possible states of the world into those satisfying and those not satisfying the goal. Though planners use goals to guide their reasoning, the crude binary distinctions defined by goals provide no basis for choosing among alternative plans that ensure achievement of goals, and no guidance whatever when no such plans can be found. These lacunae pose significant problems for planning in all realistic situations, where actions have uncertain effects or objectives can be partially satisfied. To overcome these widely-recognized expressive limitations of goals, many AI planners make ud hoc use of heuristic evaluation functions. These augment the guidance provided by goals, but lack the semantic justification needed to evaluate their true efficacy. We believe that heuristic evaluation functions should not be viewed as mere second-order refinements on the primary goal-based representation of objectives, supporting a separate “optimizing” phase of planning. Our thesis is that relative preference over the possible results of a plan constitutes the fundamental concept underlying the objectives of planning, with goals serv*Jon Doyle is supported by the USAF Rome Laboratory and DARPA under contract F30602-91-C-0018. Jon Doyle* MIT Lab for Computer Science 545 Technology Square